MATCHED FILTER BOUND OPTIMIZATION FOR MULTIUSER DOWNLINK TRANSMIT BEAMFORMING
Giuseppe Montalbano
?and Dirk T. M. Slock
??
Institut Eur´ecom
2229 Route des Crˆetes, B.P. 193, 06904 Sophia–Antipolis CEDEX, France E-Mail:
fmontalba, slock
g@eurecom.fr
Dipartimento di Elettronica, Politecnico di Torino, Torino, Italy
ABSTRACT
This paper is devoted to the optimization of the matched filter bounds (MFB) of different co-channel users, using adaptive an- tenna arrays at base stations for downlink transmit beamforming in cellular mobile communication systems. We mainly consider time division multiple access (TDMA) frequency division duplex (FDD) based systems. Note that in the case of time division du- plex (TDD), under certain assumptions the downlink channel can be assumed to be practically the same as the uplink channel. On the contrary, when using FDD, the downlink channel can not be directly observed and estimated. That makes FDD based sys- tems most difficult to deal with, although the proposed criteria and methods are general and suitable for both FDD and TDD.
Problem formulations are provided for both spatial division mul- tiple access (SDMA) and non-SDMA spectrum reuse techniques.
Novel analytical solutions and algorithms are derived, implemen- tation issues are discussed and simulations are provided in order to compare different approaches.
1. INTRODUCTION AND MOTIVATION
The use of adaptive antenna arrays at base stations allows increase the capacity of mobile radio networks by an improved spectrum efficiency, in the uplink as well as in the downlink. We investigate different approaches to optimize the weight vectors of adaptive antenna arrays at base stations for multiuser downlink transmit beamforming. We address the problem mainly in the context of time division multiple access (TDMA), frequency division duplex (FDD) based mobile communication systems. The main difficulty in transmission with antenna arrays in FDD systems consists in the limited knowledge at the base station of the downlink channel, since it can not be directly observed and therefore estimated. On the contrary, assuming the mobile velocity low enough and the re- ceiver and transmitter appropriately calibrated, the uplink and the downlink channels can be considered to be practically the same in the case of time division duplex (TDD) based systems. When con- sidering the FDD downlink, the base station needs feedback from the mobile about the downlink channel to operate in similar con- ditions as for the TDD downlink. Nevertheless such a feedback involves latency periods, due to the mobile–base station round- trip time and the processing time, resulting in a decrease of the spectrum efficiency. Moreover, in a typical outdoor propagation environment, due to the mobile velocity, such latency periods are usually not compatible with the feedback rate required to make the system reliable. This problem was addressed by Gerlach [5]
who proposed to feedback from the mobile only the information related to the downlink channel covariance matrix instead of the channel impulse response, in order to reduce the necessary feed- back rate. However, no current or conceived cellular standard is designed to support that feedback concept. On the other hand, if
such feedback is not provided, the downlink channel character- ization can only be based on the estimates of those parameters related to the uplink channel, which are relatively frequency in- dependent and of which the changing rate is slow with respect to the uplink–downlink ping-pong time. In the presence of multi- path such parameters are usually the directions of arrival (DOA), the powers and the delays associated to each path (e.g., see [2]).
Unfortunately, the phases of the paths are strongly frequency de- pendent so that they cannot be estimated from the uplink channel.
Actually in FDD mobile communication systems, in the absence of feedback, it is possible to estimate only the downlink channel covariance matrix averaged over the paths phases. That makes FDD transmission more difficult to deal with in practice, com- pared to TDD transmission. Nevertheless, since the proposed op- timization criteria assume only the knowledge of the covariance matrix of the channel between each user and each base station (eventually averaged in the time or in the frequency domain in FDD case), they apply for both FDD and TDD.
Concerning the spectrum reuse technique we consider both spatial division multiple access (SDMA) and non-SDMA. For SDMA
d
co-channel users are allocated in the same cell in the presence of one base station, assuming interference from other cells is negli- gible. For non-SDMA the same users are allocated in
d
differentcells, i.e., in the presence of
d
base stations with one user per cell.Then the optimization goal is to maximize the minimum matched filter bound (MFB) at each mobile receiver among all the
d
usersin order to improve the mobile signal quality and/or spectrum ef- ficiency.
2. GENERAL ASSUMPTIONS
We consider
d
co-channel users in both non-SDMA and SDMA.In the first case we shall consider
d
base stations, each one of them its weight vector to one user. In the second case only one base station transmits to all the users and optimizes all the corre- sponding weight vectors. In addition for SDMA, we neglect the co-channel interference. It would be possible though to mix the SDMA and non-SDMA cases.The mobile is assumed to have a single antenna element whereas each base station has an array with
m
elements. We use the term downlink transmit beamforming since only spatial combining is used at a base station. However, the emphasis here is not really on forming beams to certain directions, but rather on exploiting the diversity between multiple sensors.We assume multipath propagation but the channel not introduc- ing inter-symbol interference (ISI) (i.e., the delay spread intro- duced by the channel is less than the duration of a symbol pe- riod). When this assumption holds the MFB reduces to the signal- to-interference plus noise-ratio (SINR). When the channel intro- duces ISI, we will just optimize the SINR instead of the MFB.
3. MFB OPTIMIZATION PROBLEM FORMULATION 3.1. Non-SDMA case
LetRj iandwjdenote the covariance matrix of the channel be- tween the
j
th base station and thei
th user (including transmitter and receiver filters), and the weight vector of thej
th base station respectively. The SINR for thei
th user is given bySINR
i=
Pd wHi Riiwij=1;j 6=iwHjRj iwj
+
i (1)where
i=
2vi=
a2 and2vi, 2a represent the noise power at thei
th mobile receiver and the power of the transmitted symbols respectively. We denoteSINR
i=
ifor anyi
, throughout the paper.Hence the general optimization problem is
max
fw
i g
min
i
f
ig (2)or
min
fw
i g
max
i
f
i,1g (3)Then letwi
=
pp
iui, withkuik2= 1
, the vector of the inverse MFB’s,1= [
1,1:::
,d1]
T and the vector of the transmit powersp= [ p
1;:::; p
d]
T, whereT denotes transpose.For non-SDMA we need to limit the maximum transmit power at each base station, i.e.,kpk1
p
max. The criterion (3) can be reformulated asmin
p ;fu
i g
k ,1
k1 s.t.kpk1
p
max,kuik2= 1
8i
(4)Then we define the normalized power delivered by the
j
th basestation to the
i
th user asc
j i=
uHjRj iuj:
For any
i
we have i,1p
ic
ii=
Xj 6=i
p
jc
j i+
i:
(5)In order to account for all the users we introduce the matrixDc
= diag( c
11; :::; c
dd)
, the matrixCT defined as[
CT]
ij=
c
j i forj
6= i 0
forj = i
the vector
= [
1:::
d]
T and the matrixP= diag(
p)
. Thenwe have the following equation
,1
=
D,c1P,1[
CTp+
] :
(6)So the criterion (4) generally leads to a set of coupled problems which cannot be solved analytically.
3.2. SDMA case
In the SDMA case we have only one base station in the presence of
d
co-channel users. Then we shall replaceRj iwithRifor anyi
. So that the SINR definition (1) becomesSINR
i=
Pd wHi Riwij=1;j 6=iwHj Riwj
+
i (7)In the presence of only one base station we need to limit the over- all transmitted power to be less than or equal to
p
max. Therefore the optimization criterion ismin
p;fug k
,1
k
1 s.t.kpk1
p
max,kuik2= 1
8i
(8)Then by redefining
c
j iasc
j i=
uHjRiujwe obtain the expression of,1in the same form as (6).
Due to the constraint on the powers in the case of SDMA it can be shown that the optimum (4) leads to the same
for all the users.Indeed if the
i’s, fori = 1 ; :::; d
are not the same, then we can scale the powersfp
igto improvemin(refer to [1] for a detailed proof). On the contrary, the optimum generally does not yield the same for all the users in the case of non-SDMA because we cannot arbitrarily scale the powersp
i’s (i.e., when anyp
i= p
max no further increase ofp
iis possible).4. OPTIMIZATION IN THE ABSENCE OF NOISE In the absence of noise the SINR becomes equal to the signal-to- interference ratio (SIR) which is insensitive to the absolute power level. Then the constraint on the maximum transmit power is ir- relevant with respect to the SIR optimization in both non-SDMA and SDMA cases.
In addition we observe that the optimization problems for non- SDMA and SDMA have the same form and both of them lead to have the same SINR for all the users. So that in the sequel of this section we provide a detailed analysis only for the case of non-SDMA, and the case of SDMA will result by replacing the channel covariance matricesRj i withRi. The SIR for the
i
thuser in non-SDMA is defined as
SIR
i=
Pd wHi Riiwij=1;j 6=iwHjRj iwj (9) and the equation (6) reduces to
,1
=
D,c1P,1CTp (10)where now
i= SIR
ifor anyi
. Considering the criterion (4) and the definition (9) it is straightforward to see that the optimum is achieved when all the inter-user-interference (IUI) is zero so that i,1= 0
for alli
’s. Then the optimum approach in the absence of noise consists in forcing to zero the IUI, whenever possible.4.1. Zero-Forcing (ZF) solution
In the absence of noise at the receiver the global optimum is achieved when the following ZF conditions are satisfied
max
ku
i
k2=1uHi Riiui s.t.
X
j 6=i
p
juHjRj iuj= 0
(11)Note that the second condition in (11) is equivalent to a set of ZF conditions of the formuHi
Rijui
= 0
, i.e.,c
ij= 0
, forj
6= i
. Practically that leads to nulling the IUI while maximizing the signal-to noise ratio (SNR) at the receiver. In that caseCT=
0ddand the criterion (4) is satisfied when all the base stations transmit at the maximum power
p
max.When considering SDMA the criterion (8) is satisfied when
p
i= p
max=d
for alli
’s.Unfortunately to achieve conditions (11) we need a number of antennas greater than the sum of all the paths of all the users.
In typical mobile propagation environments the number of all the paths of all the users is usually much larger than the number of an- tennas at a base station so that ZF generally cannot be performed with purely spatial processing.
4.2. Non-ZF solutions
When the ZF conditions (11) cannot be achieved, other approaches are possible but generally the optimization cannot be carried out analytically for bothpandfuigat the same time. So we shall optimizepandfuigvia a two-step procedure eventually seeking the global optimum using an iterative algorithm.
Because of the insensitivity to the absolute value of the powers we can simplify the optimization problem by requiring each vector
u
ito have unit norm in the metricRii, i.e.,uHi R
ii u
i
= 1
forany
i
, instead ofkuik2= 1
. Under that assumption, the powerp
icorresponds to the power at thei
th receiver whereas the actual power at the related transmitter is given byp
ikuik22, for anyi
.Also,Dc
=
Idin that case.4.2.1. Power assignment optimization
We optimize the vectorpassuming a given set of direction vectors
fuig. Note that since the optimum involves
i=
for anyi
, theequation (10) reduces to
,1p=
ATp (12)whereAT
=
D,c1CT=
CT is a non-negative matrix. More- overphas to be a non-negative vector and,1 has to be non- negative as well. On the basis of the following theorems ([8],[1]) Theorem 1For a non-negative matrix, the eigenvalue of the largest norm is positive, and its corresponding eigenvector can be chosen to be non-negative.
Theorem 2
For a non-negative matrixA, the non-negative eigenvector corre- sponding to the eigenvalue of the largest norm is positive.
Theorem 3
Given the matrixAthere exists only one solution to equation (12).
we can claim that the only positive eigenvector ofAT is the one corresponding to its largest eigenvalue which is positive as well.
In other words, that ensures the existence of a positive
,1=
max(
AT)
and a unique positivep= V
max(
AT)
.4.2.2. Direction vectors optimization
Having an estimate ofp, we can optimizefuig. Indeed the opti- mization criterion is given by
min
fu
i g
max(
AT)
(13)which is equivalent to
min
fuig q
T
A T
p s.t. uHi
Riiui
= 1
(14)whereq
= V
max(
A)
. The criterion (14) leads to a set ofd
de-coupled problems whose solution is given byui
=
q eie H
i R
ii e
i
, whereei
= V
max(
Rii;
Pdj 6=iq
jRij)
for anyi
. The new set of direction vectorsfuigcan be used to re-optimize the powersp according to (12).
4.2.3. kATk1minimization based solution As initialization we can use the following criterion
min
fu
i g
kA T
k1 s.t. uHi Riiui
= 1
(15)Indeed,
max(
AT)
will be small whenAT is small. This ap- proach has the advantage of optimizing the direction vectorsfuig
independently from the powersp. It leads to a set of
d
decou-pled minimization problems whose solution is given byui
=
e
i
q
e H
i Riiei
, where, in this case,ei
= V
max(
Rii;
Pdj=1Rij)
for any
i
.Note that the criterion (15) corresponds to minimizing the power delivered to the undesired users while maximizing the power de- livered to the desired user, by each base station in non-SDMA, by each each weight vectorwi in SDMA. Apart from a differ- ent normalization ofkuik2, this criterion is equivalent to the one proposed in [3, 4].
4.2.4.
max(
AT)
minimization based algorithmAccording to the previous argumentation, we propose the iterative procedure summarized in Table 1 to find the global optimum.
Table 1:
max(
AT)
minimization based algorithm (i) Initializeuiusing (15) for anyi
;(ii) Computeq
= V
max(
A)
;(iii) Givenq, computeei
= V
max(
Rii;
Pj 6=iq
jRij)
;(iv) Computeui
=
q eie H
i R
ii e
i
; (v) Go back to (ii) until convergence;
(vi) Computewi
=
pp
iui, wherep= V
max(
AT)
;4.2.5. k,1k1minimization based algorithm
Another generally sub-optimal criterion is the following
min
p;fu
i g
k
,1k1 (16)
We remark that the optimum yieldsk,1k1
= d
k,1k1sothat near the optimum the criterion (16) is approximately equiva- lent to the criterion (4). With (16) as optimization criterion we can easily derive an iterative two-step optimization procedure, similar to the one previously described, summarized in Table 2.
Unfortunately when not near the global optimum this algorithm
Table 2:k,1k1minimization based algorithm (i) Initializeuiusing (15) for any
i
;(ii) Computep
= V
max(
AT)
;(iii) Givenp, computeei
= V
max(
Rii;
Pj 6=ip1j Rij
)
;(iv) Computeui
=
q eie H
i Riiei
; (v) Go back to (ii) until convergence;
(vi) Computewi
=
pp
iui.can converge to a local minimum, resulting in worse performance than the algorithm described in Table 1.
An optimization criterion similar to (16) was proposed in [5] as- suming a different cost function1to optimize the downlink beam- forming weights at a base station in an SDMA context. Some variants are possible to improve the performance and the numer- ical robustness of the algorithm (see [5] for further details), but convergence to the global optimum remains not guaranteed.
1In [5]irepresents the ratio between the power delivered to theith user and the interference generated to the other users by the weight vector
w
i.
5. OPTIMIZATION IN THE PRESENCE OF NOISE We showed that when the noise is present, the different constraints on the transmit powers make the optimum still leading to the same SINR for all the users in the case of SDMA but not in the case of non-SDMA. Moreover, the noise makes the optimization of the direction vectorsfuiga set of coupled problems that does not allow an analytical approach to find a solution.
Although one can observe that the criterion (16) for a givenpstill leads to the same optimization problem for the direction vectors
fu
i
gas in the noiseless case, nevertheless we shall consider that such criterion does not guarantee the convergence to the global optimum anyway. Therefore, we suggest to compute the vectors
fu
i
gby using the algorithm in Table 1 previously derived in the absence of noise and then optimize the power assignment accord- ing to the following criterion.
5.1. Power assignment optimization for SDMA
Assuming a given setfuigand all the
i’s the same, the expres- sion (6) can be arranged in order to include the constraint on the transmitted power as followsBp
~ =
,1Gp~
(17)wherep
~ = [
pT1]
T,B
=
A T
01d
0
G
=
Id 0d1
g T
,
p
max
where
=
D,c1andg= [
ku1k22:::
kudk22]
T is defined in order to havegTp= p
max. Then as in to [1] sinceGis invertible we haveEp
~ =
,1p~ ;
E=
G,1B=
2
4 A
T
g T
A T
pmax g
T
pmax
3
5 (18) which is a non-negative matrix. Hence according to theorems 1–3,
~
pcan only be the eigenvector associated to the maximum eigenvalue ofE. Further, note that we can always re-scalep~
inorder to make its last element equal to one.
5.2. Power assignment optimization for non-SDMA
In the case of non-SDMA we cannot ensure that the optimum yields the same SINR for all the users. So that the optimization problem cannot be treated via any analytical approach. As a pos- sible sub-optimal solution, given the vectorsuicomputed in the absence of noise, we setp
:
kpk1p
maxin order to satisfy the criterion (4).6. IMPLEMENTATION ISSUES
We remark that even though the computation ofuireduces to a set of decoupled problems as in the absence of noise, the computation ofpgenerally results in a coupled problem, unless ZF conditions (11) are satisfied. Hence in a non-SDMA scenario, different base stations need to communicate in order to choose the respective transmit powers. On the contrary this problem is not present in SDMA.
When the noise is present, since the base station cannot estimate the noise variance
v2i at each receiver, unless such an estimate is provided by the mobile, the vectorcannot be estimated. To remedy this drawback we shall properly define the SNR at the receiver. A possible definition is given bySNR
i= p
i imax(
Rii)
for any
i
. In practice we needmin
i
f
SNR
igSNR
min (19) whereSNR
minis a value necessary at the receiver to work with an outage probability below a specified maximum. Assuming all the users using the same receiver the worst case for thei
th useroccurs when
p
i= p
maxwhilei=
max=
kk1. Therefore a sufficient condition to satisfy the requirement (19) is given by settingSNR
min= p
max maxmin
i
f
max(
Rii)
g (20)Given
SNR
minandp
max,maxcan be derived. Then settingi=
maxfor all thei
’s the condition (19) is satisfied. Finally, note that forp
max!1the optimum solution is the one in the absence of noise, for anymax6= 0
.7. SIMULATIONS
In this section we consider an SDMA scenario in the presence of
d = 3
co-channel users in the absence of channel delay spread but in the presence of angular spread due to multipath propagation.The number of all the paths of all the users is 14 (4 for the first two users and 6 for the third, respectively). An antenna array with
m = 8
orm = 6
elements is assumed at the base station. Then ZF conditions (11) cannot be applied.Figures 1 and 2 show the convergence curves of the first and the second proposed algorithms, for
m = 8
andm = 6
respectively, in the presence of the same user scenario. As expected the sec- ond algorithm performs worse when not near the optimum. The variant proposed by Gerlach [5] in several cases makes the second algorithm perform more closely to the first one, but at the cost of an increased computational complexity and a lower convergence rate.0 1 2 3 4 5 6 7 8 9 10
20 25 30 35 40 45 50 55 60
No. of iterations
SIR (dB)
1st alg.
2nd alg.
Figure 1: Convergence curves for the two proposed algorithms,
d = 3
andm = 8
0 1 2 3 4 5 6 7 8 9 10
5 10 15 20 25
No. of iterations
SIR (dB)
1st alg.
2nd alg.
Figure 2: Convergence curves for the two proposed algorithms,
d = 3
andm = 6
10 20 30 40 50 60 70 80 90 100
−10 0 10 20 30 40 50 60
SNRmin (dB)
SINR (dB)
1st alg.
2nd alg.
2nd alg. var.
Figure 3: Optimum SINR vs. SNRminfor the 1st, the 2nd and the variant in [5] of 2nd algorithm, for
d = 3
and andm = 8
10 20 30 40 50 60 70 80 90 100
−5 0 5 10 15 20 25
SNRmin (dB)
SINR (dB)
1st alg.
2nd alg.
2nd alg. var.
Figure 4: Optimum SINR vs. SNRminfor the 1st, the 2nd and the variant in [5] of 2nd algorithm, for
d = 3
andm = 6
Finally figures 3 and 4 show the performances of the two proposed algorithms and the Gerlach’s variant of the second algorithm in the presence of noise, assuming the criterion (18) for the opti- mization of the power assignment. Once again we remark that none of the previous algorithms ensures the convergence to the global optimum in the presence of noise. That explains why the second algorithm variant (criterion (16), can perform better than the first one, as shown in figure 3.
8. RELATION TO PREVIOUS APPROACHES Some similarities among the approaches proposed here and other ones already described in the literature have already been pointed out. Several approaches [2]–[6] assumed the cost function given by the SIR at the transmitter instead of at the receiver, i.e., SIRi
was defined as the the ratio between the power delivered to the
i
th user and the interference generated to the other co-channel users by the weight vectorwi. That cost function was both op- timized user by user [2, 3, 4, 6], without assuming any multiuser power assignment optimization, leading to the same criterion as (15), and considering the multiuser power assignment optimiza- tion [5]. Other authors [7] considered the possibility of applying ZF conditions to only the dominant path of each user.A final remark concerns the power assignment optimization pre- sented in [1]. Indeed in that paper the authors considered only the spatial signatures of the users instead of the channel covariance matrices for the formulation of the problem (18). Hence, in the presence of channel delay spread the formulation in [1] is not ap- propriate and does not lead to the optimization of the SINR of all the users.
9. CONCLUSIONS
In this paper we addressed the problem of the optimization of the MFB with respect to the beamforming weights at base sta-
tions, for multiuser downlink transmission in both non-SDMA and SDMA spectrum reuse techniques. A general problem for- mulation yielded the definition of a proper cost function to be minimized. Then we considered the optimization problem in the absence of noise. In that case the ZF solution represents the opti- mum, but it can be achieved only under certain conditions usually not verified in practice. Therefore a novel algorithm to find the global optimum in the absence of noise has been derived by an an- alytical approach which does not require the conditions necessary to the ZF solution. The novel algorithm has shown to outperform other sub-optimal algorithms based on different optimization cri- teria, in terms of optimum MFB, convergence rate, numerical ro- bustness and computational complexity. We remarked that the latter sub-optimal are very similar to other ones already proposed in the literature [3, 4, 5], generally derived starting from different cost functions.
The optimization problem in the presence of noise has also been addressed, considering the constraints on the transmit power in- herent to both SDMA and non-SDMA spectrum reuse techniques.
At present, we are not aware of any analytical approach to find a global optimum in that case and we are currently investigating non-analytical optimization approaches. Only the assignment of the transmitted powers [1] can be easily optimized in the SDMA case for a given set of direction vectors. So that for that aim we propose to use the set of direction vectors optimized in the ab- sence of noise through the algorithm we described in Table 1. Fi- nally we analyzed via simulation how optimum and sub-optimum algorithms derived in the absence of noise, perform in the pres- ence of noise with the appropriate power assignment optimiza- tion.
10. REFERENCES
[1] W. Yang and G. Xu, “Optimal downlink power assignment for smart antenna systems,” Proc. ICASSP‘98, Vol. 5, pp.
3337–3340, May 1998.
[2] G. G. Raleigh and V. K. Jones, “Adaptive antenna transmis- sion for frequency duplex digital wireless communication,”
Proc. ICC‘97, Vol. 2, pp. 641–646, June 1997.
[3] P. Zetterberg Mobile cellular communications with base sta- tion antenna arrays: Spectrum efficiency, algorithms and propagation models, Ph.D. thesis, Royal Inst. of Technol- ogy, Stockholm, Sweden, 1997.
[4] J. Goldberg and J. Fonollosa, “Downlink beamforming for cellular mobile communications,” Proc. VTC‘97, Vol. 2, pp.
632–636, May 1997.
[5] D. Gerlach, Adaptive transmitting antenna arrays at the base station in mobile radio network, Ph.D. Thesis, ISL, Stanford University, Stanford, USA, 1995.
[6] G. G. Raleigh, S. N. Diggavi, V. K. Jones, and A. Paulraj,
“A blind transmit antenna algorithms for wireless commu- nication,” Proc. ICC‘95, Vol. 2, pp. 1494–1499, June 1995.
[7] G. Xu and H. Liu, “An effective transmission beamforming scheme for frequency-division-duplex digital wireless com- munication systems,” Proc. ICASSP‘95, pp. 1729–1732, 1995.
[8] R. A. Horn and C. R. Johnson, Matrix analysis, Cambridge University Press, 1985.